File size: 13,101 Bytes
629b123 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | # SPDX-FileCopyrightText: © 2025 Tenstorrent AI ULC
# SPDX-License-Identifier: Apache-2.0
"""Anti-aliased 1D resampling layers for LTX-2 audio (BigVGAN-v2 style).
Device equivalents of the reference's ``LowPassFilter1d`` / ``UpSample1d`` /
``DownSample1d`` / ``Activation1d``. The transposed-conv upsample is expressed as
zero-stuff + zero-pad + a depthwise tap filter (bit-equivalent to
``F.conv_transpose1d`` for the symmetric sinc kernel). ``UpSample1d`` is shared
by the vocoder (kaiser window, T-shard/halo aware) and the BWE resampler (hann
window, single-device).
"""
from __future__ import annotations
import math
import torch
import ttnn
from ..parallel.config import ParallelFactor
from ..parallel.manager import CCLManager
from .audio_ops import (
_make_kaiser_sinc_kernel_1d,
_replicate_pad_t,
_t_neighbor_pad,
_zero_pad_t,
_zero_stuff_t,
depthwise_tap_filter,
)
from .module import Module
def _make_hann_sinc_kernel_1d(*, ratio: int) -> tuple[torch.Tensor, int, int, int, int]:
"""Return ``(kernel, kernel_size, pad, pad_left_crop, pad_right_crop)`` for the
Hann-window sinc resampler (torchaudio-equivalent), used by the BWE skip path."""
rolloff = 0.99
lowpass_filter_width = 6
width = math.ceil(lowpass_filter_width / rolloff)
kernel_size = 2 * width * ratio + 1
pad = width
pad_left = 2 * width * ratio
pad_right = kernel_size - ratio
time_axis = (torch.arange(kernel_size, dtype=torch.float64) / ratio - width) * rolloff
time_clamped = time_axis.clamp(-lowpass_filter_width, lowpass_filter_width)
window = torch.cos(time_clamped * math.pi / lowpass_filter_width / 2) ** 2
sinc_filter = torch.sinc(time_axis) * window * rolloff / ratio
return sinc_filter.float().reshape(kernel_size), kernel_size, pad, pad_left, pad_right
class LowPassFilter1d(Module):
"""Depthwise low-pass conv1d with a fixed kaiser-sinc kernel.
The kernel is constant (baked at __init__), so ``_prepare_torch_state`` only
absorbs a checkpoint-provided kernel if present (BigVGAN convention).
"""
def __init__(
self,
*,
cutoff: float = 0.5,
half_width: float = 0.6,
stride: int = 1,
kernel_size: int = 12,
padding: bool = True,
padding_mode: str = "replicate",
mesh_device: ttnn.MeshDevice,
dtype: ttnn.DataType = ttnn.float32,
parallel_config: ParallelFactor | None = None,
ccl_manager: CCLManager | None = None,
) -> None:
super().__init__()
if cutoff < 0.0 or cutoff > 0.5:
raise ValueError("cutoff must be in [0, 0.5]")
if padding_mode not in ("replicate", "zeros"):
raise ValueError(f"padding_mode must be replicate or zeros, got {padding_mode!r}")
sharded = parallel_config is not None and parallel_config.factor > 1
if sharded:
assert ccl_manager is not None, "T-sharding requires ccl_manager"
self.kernel_size = kernel_size
self.even = kernel_size % 2 == 0
self.pad_left = kernel_size // 2 - int(self.even)
self.pad_right = kernel_size // 2
self.stride = stride
self.padding = padding
self.padding_mode = padding_mode
self.mesh_device = mesh_device
self.dtype = dtype
self.parallel_config = parallel_config
self.ccl_manager = ccl_manager
kernel = _make_kaiser_sinc_kernel_1d(cutoff, half_width, kernel_size)
self._taps_cpu = kernel.tolist()
self._conv1d_cache: dict = {}
def _prepare_torch_state(self, state: dict[str, torch.Tensor]) -> None:
if "filter" in state:
t = state.pop("filter")
assert tuple(t.shape) == (1, 1, self.kernel_size)
self._taps_cpu = t.reshape(self.kernel_size).float().tolist()
def forward(self, x_BTC: ttnn.Tensor) -> ttnn.Tensor:
"""``x_BTC``: ``(B, T, C)`` ROW_MAJOR. Returns ``(B, T_out, C)``.
When ``parallel_config.factor > 1``, ``T`` is the *per-device* extent and
the replicate/zero pad becomes a halo exchange via ``_t_neighbor_pad``.
"""
assert x_BTC.layout == ttnn.ROW_MAJOR_LAYOUT
sharded = self.parallel_config is not None and self.parallel_config.factor > 1
if self.padding:
if sharded:
x = _t_neighbor_pad(
x_BTC,
pad_left=self.pad_left,
pad_right=self.pad_right,
parallel_config=self.parallel_config,
ccl_manager=self.ccl_manager,
padding_mode=self.padding_mode,
)
elif self.padding_mode == "replicate":
x = _replicate_pad_t(x_BTC, self.pad_left, self.pad_right, self.mesh_device)
else:
x = _zero_pad_t(x_BTC, self.pad_left, self.pad_right, self.mesh_device)
else:
x = x_BTC
return depthwise_tap_filter(
x,
self._taps_cpu,
self.stride,
mesh_device=self.mesh_device,
dtype=self.dtype,
cache=self._conv1d_cache,
)
class UpSample1d(Module):
"""Anti-aliased sinc upsampler (zero-stuff + depthwise lowpass).
``window="kaiser"`` is the BigVGAN anti-alias upsampler (T-shard/halo aware);
``window="hann"`` is the torchaudio-equivalent resampler for the BWE skip path
(single-device). The depthwise Conv1d formulation is bit-equivalent to
``F.conv_transpose1d`` for the symmetric sinc kernel.
"""
def __init__(
self,
*,
ratio: int = 2,
window: str = "kaiser",
kernel_size: int | None = None,
mesh_device: ttnn.MeshDevice,
dtype: ttnn.DataType = ttnn.float32,
parallel_config: ParallelFactor | None = None,
ccl_manager: CCLManager | None = None,
) -> None:
super().__init__()
sharded = parallel_config is not None and parallel_config.factor > 1
if sharded:
assert ccl_manager is not None, "T-sharding requires ccl_manager"
self.ratio = ratio
self.stride = ratio
self.window = window
self.mesh_device = mesh_device
self.dtype = dtype
self.parallel_config = parallel_config
self.ccl_manager = ccl_manager
if window == "hann":
kernel, self.kernel_size, self.pad, self.pad_left_crop, self.pad_right_crop = _make_hann_sinc_kernel_1d(
ratio=ratio
)
self._taps_cpu = kernel.tolist()
elif window == "kaiser":
self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
self.pad = self.kernel_size // ratio - 1
self.pad_left_crop = self.pad * self.stride + (self.kernel_size - self.stride) // 2
self.pad_right_crop = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
kernel = _make_kaiser_sinc_kernel_1d(
cutoff=0.5 / ratio, half_width=0.6 / ratio, kernel_size=self.kernel_size
)
self._taps_cpu = kernel.tolist()
else:
raise ValueError(f"window must be kaiser or hann, got {window!r}")
self._conv1d_cache: dict = {}
self._use_polyphase = (self.kernel_size % ratio) == 0
if self._use_polyphase:
self._poly_K_sub = self.kernel_size // ratio
def _prepare_torch_state(self, state: dict[str, torch.Tensor]) -> None:
if "filter" in state:
t = state.pop("filter")
assert tuple(t.shape) == (1, 1, self.kernel_size)
self._taps_cpu = t.reshape(self.kernel_size).float().tolist()
def forward(self, x_BTC: ttnn.Tensor) -> ttnn.Tensor:
assert x_BTC.layout == ttnn.ROW_MAJOR_LAYOUT
B, T, C = x_BTC.shape
sharded = self.parallel_config is not None and self.parallel_config.factor > 1
poly2 = self._use_polyphase and self.ratio == 2
# ratio-2 polyphase reads only x_pad[2:T_pad-2]; pad two fewer rows per side when possible.
crop = 2 if (poly2 and self.pad >= 2) else 0
eff_pad = self.pad - crop
if sharded and eff_pad > 0:
x_pad = _t_neighbor_pad(
x_BTC,
pad_left=eff_pad,
pad_right=eff_pad,
parallel_config=self.parallel_config,
ccl_manager=self.ccl_manager,
padding_mode="replicate",
)
else:
x_pad = _replicate_pad_t(x_BTC, eff_pad, eff_pad, self.mesh_device)
if poly2:
B_, T_pad, C_ = x_pad.shape
# Zero-pad sub-taps (sub0 trailing, sub1 leading) so both phases convolve the same input.
if crop:
base = x_pad
else:
base = ttnn.slice(x_pad, [0, 2, 0], [B_, T_pad - 2, C_])
scaled_taps = [t * self.ratio for t in self._taps_cpu]
sub0 = [scaled_taps[2 * j + 0] for j in range(self._poly_K_sub)] + [0.0]
sub1 = [0.0] + [scaled_taps[2 * j + 1] for j in range(self._poly_K_sub)]
ph0 = depthwise_tap_filter(
base,
sub0,
1,
mesh_device=self.mesh_device,
dtype=self.dtype,
cache=self._conv1d_cache,
)
ph1 = depthwise_tap_filter(
base,
sub1,
1,
mesh_device=self.mesh_device,
dtype=self.dtype,
cache=self._conv1d_cache,
)
if base is not x_pad:
ttnn.deallocate(base)
T_out = ph0.shape[1]
ph0_b = ttnn.reshape(ph0, (B_, T_out, 1, C_))
ph1_b = ttnn.reshape(ph1, (B_, T_out, 1, C_))
stacked = ttnn.concat([ph0_b, ph1_b], dim=2)
return ttnn.reshape(stacked, (B_, T_out * 2, C_))
x_zs = _zero_stuff_t(x_pad, stride=self.stride, mesh_device=self.mesh_device)
x_padded = _zero_pad_t(x_zs, self.kernel_size - 1, self.kernel_size - 1, self.mesh_device)
y = depthwise_tap_filter(
x_padded,
[t * self.ratio for t in self._taps_cpu],
1,
mesh_device=self.mesh_device,
dtype=self.dtype,
cache=self._conv1d_cache,
)
T_y = y.shape[1]
y_cropped = ttnn.slice(y, [0, self.pad_left_crop, 0], [B, T_y - self.pad_right_crop, C])
ttnn.deallocate(y)
return y_cropped
class DownSample1d(Module):
"""Strided kaiser-sinc lowpass downsampler wrapping ``LowPassFilter1d``."""
def __init__(
self,
*,
ratio: int = 2,
kernel_size: int | None = None,
mesh_device: ttnn.MeshDevice,
dtype: ttnn.DataType = ttnn.float32,
parallel_config: ParallelFactor | None = None,
ccl_manager: CCLManager | None = None,
) -> None:
super().__init__()
self.ratio = ratio
self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
self.lowpass = LowPassFilter1d(
cutoff=0.5 / ratio,
half_width=0.6 / ratio,
stride=ratio,
kernel_size=self.kernel_size,
padding=True,
padding_mode="replicate",
mesh_device=mesh_device,
dtype=dtype,
parallel_config=parallel_config,
ccl_manager=ccl_manager,
)
def forward(self, x_BTC: ttnn.Tensor) -> ttnn.Tensor:
return self.lowpass(x_BTC)
class Activation1d(Module):
"""Anti-aliased activation: ``UpSample1d(2x) → activation → DownSample1d(2x)``."""
def __init__(
self,
*,
channels: int,
activation: Module,
up_ratio: int = 2,
down_ratio: int = 2,
up_kernel_size: int = 12,
down_kernel_size: int = 12,
mesh_device: ttnn.MeshDevice,
dtype: ttnn.DataType = ttnn.float32,
parallel_config: ParallelFactor | None = None,
ccl_manager: CCLManager | None = None,
) -> None:
super().__init__()
self.channels = channels
self.act = activation
self.upsample = UpSample1d(
ratio=up_ratio,
window="kaiser",
kernel_size=up_kernel_size,
mesh_device=mesh_device,
dtype=dtype,
parallel_config=parallel_config,
ccl_manager=ccl_manager,
)
self.downsample = DownSample1d(
ratio=down_ratio,
kernel_size=down_kernel_size,
mesh_device=mesh_device,
dtype=dtype,
parallel_config=parallel_config,
ccl_manager=ccl_manager,
)
def forward(self, x_BTC: ttnn.Tensor) -> ttnn.Tensor:
y = self.upsample(x_BTC)
y = self.act(y)
if y.layout != ttnn.ROW_MAJOR_LAYOUT:
y = ttnn.to_layout(y, ttnn.ROW_MAJOR_LAYOUT)
y = self.downsample(y)
return y
|